The gambler’s fallacy can apply the other way, e.g. “This die has rolled over 3 many times, so it will probably roll 3 or lower next time.” So that's not really the issue.
Your dice example is not illustrating the point the GP is trying to make. It’s just comparing the probabilities of two outcomes of the same distribution (the sum-of-two-dice distribution).
This has nothing to do with the selection effect the GP is talking about. That would be recording the value of the die that rolled highest (i.e. taking the maximum over dice) and then focusing on the same die next time (i.e. not taking the maximum over dice).
It’s comparing the probability distribution of max {X_1, X_2, X_3, ...} with the probability distribution of any individual X_i. If A is sampled from the former and B is sampled from the latter, P(B < A) > P(B > A).
With regard to the race, if you do really well, much better than your average, without any fundamental change to your ability that's the same as rolling a 12. It's an unlikely (though possible) event that happened to occur. Over a series of races, however, we would not anticipate a repeat of that performance (that is, it's a low probability event, and a series of such low probability events has even lower probability).
The gambler's fallacy goes the other way. It's the belief that if we've had a string of bad luck [0], then we should wager on good luck being around the corner. That is, the gambler assumes a low probability event is "inevitable" after a series of high probability events. Regression to the mean: A high probability event should be anticipated after a low probability event.
[0] In gambling this usually means getting close to the average, that is: the average hand is the lowest valued hand, the rarest hand is the highest valued. So a gambler making this fallacy would wager on seeing a royal flush after several hundred hands of only seeing pairs.